Multivariable virtual diagnostics and tuning of beam positioning using machine learning

Abstract

In many experiments, such as Thomson scattering experiments, beam position sensitivity is encountered when using the Tsinghua Thomson Scattering X-ray Source (TTX). In each experiment, the position of the beam is manually tuned, a process that requires several minutes to complete. Moreover, the beam position drifts over time. In addition, in some situations, such as free electron laser operation, the beam position will be affected; thus, a downstream beam position monitor (BPM) cannot offer a reliable beam position. In this study, machine learning is used to perform virtual diagnostics and tuning of the beam position of TTX. For universality, 20 independent parameters that may affect the beam position are imported and tuned over a large range. While universal virtual diagnostics can be achieved using this process, the parameters mutually influence each other, and it is difficult to determine their valid combined range, leading to low data validity; consequently, in some situations, the beam will be lost.A bagged tree classification model is trained to determine the combined range of the parameters. This greatly improves the experimental efficiency from 38% to 64%. The final accuracy achieved using this classification model is 95%. When the beam perturbations have a root mean square error (RMSE) of approximately 0.096 mm, the virtual diagnostic accuracy is 0.13 mm in the range of 20 mm; thus, the normalized root mean square error (NRMSE) is 0.65%. This can satisfy most demands on TTX. In addition, it is experimentally demonstrated that the beam can be tuned to the desired position with an RMSE of 0.03 mm using this model.